Current Server Time: 18:45 (Central Europe)

#apertus IRC Channel Logs

2016/01/11

Timezone: UTC


23:47
wakalixes
left the channel
01:02
Bertl
off to bed now ...
01:02
Bertl
changed nick to: Bertl_zZ
01:25
pozitrono
joined the channel
03:31
troy_s
irieger: That is what I said; dual greens. Impossible to do two readouts in succession without a temporal bit of strangeness I believe?
03:31
troy_s
irieger: Offering simultaneous identical green filters but at two different simultaneous gains would account for solid colour and greater latitude.
03:32
troy_s
(Lean on the other green to figure out proper range when out of zone.)
05:13
jucar
joined the channel
05:42
jucar
left the channel
06:24
alexML
troy_s: they can do two readouts at the same time, with two different amplifiers on the same signal
06:24
alexML
imagine you have two audio amps connected to the same music source
06:24
alexML
you turn one louder and the other one less loud
06:25
alexML
two greens may give a little more latitude, but what do you do when the other two channel clip?
06:25
alexML
07:42 -!- jucar [~jucar@59.90.167.251] has quit [Ping timeout: 276 seconds]
06:25
alexML
08:24 < alexML> troy_s: they can do two readouts at the same time, with two different amplifiers on the same signal
06:26
alexML
07:42 -!- jucar [~jucar@59.90.167.251] has quit [Ping timeout: 276 seconds]
06:26
alexML
08:24 < alexML> troy_s: they can do two readouts at the same time, with two different amplifiers on the same signal
06:26
alexML
(sorry, that was my kid)
07:03
se6astian|away
changed nick to: se6astian
07:06
se6astian
good morning
07:06
slikdigit
left the channel
07:13
arpu
joined the channel
07:42
Bertl_zZ
changed nick to: Bertl
07:42
Bertl
morning folks!
07:43
Bertl
alexML: I like your kid ... how old is (s)he?
07:53
alexML
one year old
07:54
Bertl
it's always good to start early with IRC :)
07:56
Bertl
I got your "log" proposal for 4k encoding (from sebastian)
07:57
alexML
yeah, I remember we talked about it a while ago
07:57
Bertl
do I understand that right that it is based on YCbCr out instead of RGB?
07:57
alexML
that one is RGB based
07:58
alexML
I also tried a YCbCr version, but didn't get better results
07:58
alexML
so, the simplest one would be to apply a curve, and then, direct mapping to RGB
07:58
Bertl
with alternating greens
07:59
alexML
yep
07:59
Bertl
and yes, I get that the math is on RGB, but your reconstruction is based on YCbCr 422
07:59
Bertl
btw, it should be already implemented in the current setup
08:00
alexML
from the recorder, you get Y422 and then you convert to RGB?
08:00
Bertl
because we have a gamma LUT for HDMI out, and the alternating green should be there as well
08:01
Bertl
the only thing we need to do is construct a proper lookup table (or four of them to be precise)
08:01
alexML
in my simulations I used the same LUT for all channels
08:02
alexML
so, as a starting point, should be worth trying, I think
08:02
Bertl
I wonder if it wouldn't make sense to use a slightly shifted lut for alternating frames (in the future, not possible at the moment)
08:04
alexML
to get a little more bit depth? yes, makes sense
08:07
Bertl
another option would be to use alternating pixels for Cb and Cr to compensate for the 422
08:10
Bertl
i.e. probably would (r,g1,b), (b,g2,r) give better results in reconstruction
08:11
pozitrono
left the channel
08:13
niemand
joined the channel
08:15
alexML
worth trying, though introduces more flicker[D[D[D[D[D[D[D[D[D[D[D[D[D[D[D[D[D[D[D[D[D[D[Dit
08:15
cbohnens|away
changed nick to: cbohnens
08:20
Bertl
flicker is not a problem I thin
08:20
Bertl
\+k
08:21
Bertl
the codecs are all intra frame codecs, so no adverse effect on the recording
08:39
se6astian
plus we have two hdmi slots on the Beta so we can use one for recording and one for previewing
08:39
se6astian
of course if we can do both over the same thats even better
08:43
se6astian
anyone ever heard of https://www.mollie.com/en/pricing ?
08:43
se6astian
they also do credit card and bitcoin processing
08:43
se6astian
good rates, but never heard of them before
08:53
alexML
btw, 422 averagees 2 pixels horizontally, so we may consider doubling the horizontal resolution in some channels (transmit left half, then right half), or shifting some channels horizontally by 1 pixel
09:01
Bertl
does it actually average them?
09:02
Bertl
isn't the spacial relation fixed in YCbCr and the averaging only happens at reconstruction?
09:02
Bertl
(should be easy to test with an artificial test pattern)
09:05
alexML
well, each group of 2 pixels have two luma values and one shared chroma (Cb+Cr)
09:05
alexML
so, chroma is probably averaged
09:05
alexML
(or otherwise combined)
09:05
Bertl
that's the question, it could also be Cb from the first pixel and Cr from the second
09:06
Bertl
or Cb/Cr from every second pixel
09:06
alexML
yeah, but I think it's unlikely, since that may introduce aliasing
09:06
Bertl
we should do a simple color grid test
09:07
alexML
it might be as well a slightly larger filter
09:08
Bertl
how about a single pixel red/blue bar every 7 or 13 pixels?
09:08
Bertl
and the inverse at the bottom half or so>
09:09
Bertl
s/>/?
09:10
alexML
as a test pattern? yes
09:11
alexML
there may be even helpful patterns in what you already recorded (didn't manage to look into them yet, only decoded the memory dumps and recovered one image)
10:15
niemand
left the channel
10:46
jucar
joined the channel
10:56
jucar
left the channel
11:35
baldand
left the channel
11:35
baldand
joined the channel
12:19
niemand
joined the channel
14:14
arpu
left the channel
14:48
niemand
left the channel
14:50
cbohnens
changed nick to: cbohnens|away
15:00
se6astian
changed nick to: se6astian|away
15:00
pozitron
joined the channel
15:08
Bertl
left the channel
15:40
arpu
joined the channel
16:02
se6astian|away
changed nick to: se6astian
16:03
alexML
I think I just pulled 2 extra stops of highlights out of the hat
16:03
se6astian
hoho!
16:03
alexML
( intracube knows how - from linearizing what troy_s called "nondata" )
16:04
alexML
https://dl.dropboxusercontent.com/u/4124919/bleeding-edge/apertus/highlights/it8-gainx1-offset2047-80ms-02.jpg
16:04
alexML
https://dl.dropboxusercontent.com/u/4124919/bleeding-edge/apertus/highlights/it8-gainx1-offset2047-20ms-02.jpg
16:04
alexML
https://dl.dropboxusercontent.com/u/4124919/bleeding-edge/apertus/highlights/HTC-gainx1-offset2047-80ms-02.jpg
16:04
alexML
https://dl.dropboxusercontent.com/u/4124919/bleeding-edge/apertus/highlights/HTC-gainx1-offset2047-20ms-02.jpg
16:05
alexML
the 20ms it8 was exposed to the right (the 25ms one was already clipping)
16:05
alexML
so, I got a set of curves that matched the 80ms it8 to the 20ms one
16:05
alexML
then, used the same curves to HTC 80ms
16:05
alexML
so, IT8 is training data and HTC is validation data
16:07
alexML
besides LUTs, I used dark frame, clip frame (nonlinearity-gainx1-offset2047-75ms-01.raw12), and extreme row noise correction
16:08
alexML
the recovered highlights are a little noisy, but should be better than solid white
16:15
jucar
joined the channel
16:30
Bertl
joined the channel
16:47
troy_s
alexML: It _is_ non data.
16:47
troy_s
You don't have to trust me on that, but do the colorimetry. It is junk.
16:47
troy_s
Absolutely fine as a post production pass after all the work is donr
16:47
troy_s
But as "data" nothing further from the truth.
16:49
irieger
troy_s: Would sign that. That data isn't usable for serious work.
16:51
davidak
joined the channel
16:56
Bertl
off for a nap ... bbl
16:56
Bertl
changed nick to: Bertl_zZ
17:51
troy_s
irieger: Amen.
17:52
troy_s
irieger: again, perfectly acceptable as a planned oopsie fix in the visual effects domain (hello Zodiac ;) ) but not for the steel and wood of structure.
17:56
niemand
joined the channel
18:14
slikdigit
joined the channel
18:15
slikdigit
left the channel
18:15
slikdigit
joined the channel
18:27
intracube
alexML: nice :)
18:27
intracube
the only caveat being how adversely affected the colour accuracy is
18:29
alexML
I'd rather say the repeatability
18:30
alexML
if that one is good, you can tweak the curves to get good colors; otherwise, you'd better desaturate that :P
18:31
intracube
https://dl.dropboxusercontent.com/u/4124919/bleeding-edge/apertus/highlights/it8-gainx1-offset2047-80ms-02.jpg
18:32
intracube
there's some pink highlights in the bottom right and green edge to the shadow
18:32
intracube
I guess this is the limits of the colour when pushed this far?
18:33
alexML
that was because my "clip frame" was not actually clipped in that corner
18:34
alexML
I should have taken a slightly more overexposed image
18:34
intracube
ah ok
18:34
alexML
this is just the first proof of concept, without any claims of perfect colors
18:35
intracube
:)
18:35
alexML
(so, a little early to jump in and call it junk, imo)
18:36
intracube
didn't mean to at all
18:36
alexML
not you :P
18:41
intracube
but even in worst case, it could save the need for an expensive re-shoot
18:50
alexML
especially since this sensor barely has 9-10 stops of DR
18:51
alexML
sure, you can always use PLR instead of this, if the temporal artifacts aren't bad
18:54
alexML
(just commited the code, btw)
19:22
MartinS
joined the channel
19:47
jucar
left the channel
19:52
jucar
joined the channel
19:56
pozitron
left the channel
19:58
philippej_
joined the channel
19:59
philippej_
is always amazed to read the irc log and discover what collective intelligence grought to the table since last visit. Congrats to everyone !
19:59
philippej_
(brought:-) )
20:39
intracube
doesn't the piece-wise linear mode avoid temporal problems?
20:46
se6astian
it has its own temporal problems
20:46
se6astian
the exposure is interupted up to two times
20:56
intracube
ah ok.
20:57
intracube
will be interesting to see what that looks like in practice
21:07
jucar
left the channel
21:07
niemand
left the channel
21:15
philippej_
left the channel
22:04
se6astian
off to bed
22:04
se6astian
changed nick to: se6astian|away
22:54
pozitrono
joined the channel